ClearML/clearml-docs
1
0
Fork 0
mirror of https://github.com/clearml/clearml-docs synced 2025-05-15 09:56:15 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Clarify pipelines (#10)

Browse Source
This commit is contained in:
pollfly 2021-07-14 10:39:45 +03:00 committed by GitHub
parent 13b7ac4d82
commit 5662c42a64
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 50 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

105
docs/fundamentals/pipelines.md
View File

@ -2,73 +2,68 @@
title: Pipelines title: Pipelines
--- ---
Users can automate [Tasks](task) to run consecutively or according to some logic by putting the Tasks into a pipeline. Users can automate [Tasks](task) to run consecutively or according to some logic by putting the tasks into a pipeline.
Tasks in a pipeline can leverage other tasks' work products such as artifacts and parameters. Tasks in a pipeline can leverage other tasks' work products such as artifacts and parameters.
Pipelines are controlled by a *Controller Task* that holds the logic of the pipeline execution steps. Pipelines are controlled by a *Controller Task* that holds the logic of the pipeline execution steps.
## How do pipelines work? ## How do pipelines work?
Before running a pipeline, we need to configure a Controller Task, in which the pipeline is defined. The user decides the controlling logic, whether it be simple Before running a pipeline, we need to configure a Controller Task, in which the pipeline is defined. Pipelines are made
([DAG](https://en.wikipedia.org/wiki/Directed_acyclic_graph)) or complex custom logic. up of steps. Each step consists of a task that already exists in the ClearML Server and is used as a template. The
user decides the controlling logic of the step interactions, whether it be simple ([DAG](https://en.wikipedia.org/wiki/Directed_acyclic_graph))
or more complex.
Once the pipeline is running, it first clones existing Tasks (called templates) and then sends the cloned Tasks for execution Once the pipeline is running, it starts sequentially launching the steps configured in the Controller. In each step, the template task
according to the pipeline's control logic. is cloned, and the cloned task is sent for execution. Depending on the specifications laid out in the Controller Task, a
step's parameters can be overridden, and / or a step can use a previous step's work products.
Callbacks can be utilized to control pipeline execution flow. A callback can be defined
to be called before and / or after the execution of every task in a pipeline. Additionally, there is an option to
create customized, step-specific callbacks.
![Pipeline chart](../img/fundamentals_pipeline.png)
![image](../img/fundamentals_pipeline.png)
## Simple DAG Pipelines ## Simple DAG Pipelines
For a simple, DAG based logic, use the off-the-shelf `PipelineController` class to define the DAG (see an example [here](../guides/pipeline/pipeline_controller)). Once PipelineController object is populated and configured, For a simple, DAG based logic, use the off-the-shelf [`PipelineController`](../references/sdk/automation_controller_pipelinecontroller.md) class to define the DAG (see an example
we can start the pipeline, which will launch its first steps, then it waits until the pipeline is completed. [here](../guides/pipeline/pipeline_controller)). Once the `PipelineController` object is populated and configured,
we can start the pipeline, which will begin executing the steps in succession, then it waits until the pipeline is completed.
The pipeline control logic is processed in a background thread. The pipeline control logic is processed in a background thread.
## Custom Pipelines
In cases where a DAG is insufficient (for example when needing to launch one pipeline, then, if performance is inadequate, rerun pipeline again),
users can apply custom logic, using a generic methods to enqueue Tasks, implemented in python code.
The logic of the pipeline sits in a *Controller Task*.
Since a pipeline *Controller Task* is a Task on its own, it's possible to have pipelines running other pipelines.
This gives users greater degrees of freedom for automation.
Custom pipelines usually involves cloning existing Tasks (Template Tasks), modiftying their parameters and manually enqueuing
them to queues (For execution by [agents](../clearml_agent.md). Since it's possible to control Task's execution (Including
overriding Hyperparameters and Artifacts) and get output metrics, it's possible to create custom logic that controls inputs and acts upon outputs.
A simple Custom pipeline may look like this:
```python
task = Task.init('examples', 'Simple Controller Task', task_type=Task.TaskTypes.controller)
# Get a reference to the task to pipe to.
first_task = Task.get_task(project_name='PROJECT NAME', task_name='TASK NAME')
# Clone the task to pipe to. This creates a task with status Draft whose parameters can be modified.
cloned_first_task = Task.clone(source_task=first_task, name='Auto generated cloned task')
cloned_first_task.set_parameters({'key':val})
Task.enqueue(cloned_first_task.id, queue_name='QUEUE NAME')
# Here comes custom logic
#
#
###
# Get a reference to the task to pipe to.
next_task = Task.get_task(project_name='SECOND PROJECT NAME', task_name='SECOND TASK NAME')
# Clone the task to pipe to. This creates a task with status Draft whose parameters can be modified.
cloned_task = Task.clone(source_task=next_task, name='Second Cloned Task')
Task.enqueue(cloned_task.id, queue_name='QUEUE NAME')
```
See an example for custom pipelines [here](https://github.com/allegroai/clearml/tree/master/examples/automation)
:::note :::note
We recommend enqueuing Pipeline Controller Tasks into a We recommend enqueuing Pipeline Controller Tasks into a
[services](agents_and_queues#services-agent--queue) queue [services](agents_and_queues.md#services-agent--queue) queue
::: :::
Callback functions can be specified to be called in the steps of a `PipelineController` object.
There is an option to define a callback to be called before and / or after every step in the pipeline,
using the `step_task_created_callback` or the `step_task_completed_callback` parameters of the [`start`](../references/sdk/automation_controller_pipelinecontroller.md#start)
method. Alternatively, step-specific callback functions can be specified with the `pre_execute_callback` and / or
`post_execute_callback` parameters of the [`add_step`](../references/sdk/automation_controller_pipelinecontroller.md#add_step)
method.
## Advanced pipelines
Since a pipeline *Controller Task* is itself a ClearML Task, it can be used as a pipeline step and can be used to create
more complicated workflows, such as pipelines running other pipelines, or a pipeline running multiple tasks concurrently.
For example, it could be useful to have one pipeline for data preparation, which triggers a second pipeline that trains
networks.
It could also be useful to run a pipeline that runs tasks concurrently, training multiple networks with different hyperparameter
values simultaneously. See the [Tabular training pipeline](../guides/frameworks/pytorch/notebooks/table/tabular_training_pipeline.md)
example of a pipeline with concurrent steps.
## Custom Pipelines
In cases where a DAG is insufficient (for example, when needing to launch one pipeline, then, if performance is inadequate,
rerun pipeline again), users can apply custom logic, using generic methods to enqueue tasks, implemented in python code.
The custom logic of the pipeline sits in the *Controller Task*'s script.
Custom pipelines usually involve cloning template tasks, modifying their parameters, and manually enqueuing
them to queues (for execution by [agents](../clearml_agent.md)). It's possible to create custom logic that controls inputs
(e.g. overriding hyperparameters and artifacts) and acts upon task outputs.
See an example of a custom pipeline [here](../guides/automation/task_piping.md).